1

ITK Lecture 13:Registration

ITK Lecture 13:Registration

Damion Shelton

Methods in Image Analysis

CMU Robotics Institute 16-725

U. Pitt Bioengineering 2630

Spring Term, 2006

Damion Shelton

Methods in Image Analysis

CMU Robotics Institute 16-725

U. Pitt Bioengineering 2630

Spring Term, 2006

2

For more info/gory detail…For more info/gory detail…

Please see Chapter 8 in the Software Guide and Insight into Images for exhaustive detail

Please see Chapter 8 in the Software Guide and Insight into Images for exhaustive detail

3

What is registration?What is registration?

The process of aligning a target image to a source image

More generally, determining the transform that maps points in the target image to points in the source image

The process of aligning a target image to a source image

More generally, determining the transform that maps points in the target image to points in the source image

4

Transform typesTransform types

Rigid (rotate, translate) Affine (rigid + scale & shear) Deformable (affine + curvature) Many others

Rigid (rotate, translate) Affine (rigid + scale & shear) Deformable (affine + curvature) Many others

5

Registration in ITKRegistration in ITK

ITK uses an extensible registration framework

Various interchangeable classes exist Relatively easy to “twiddle” the part

you’re interested in while recycling prior work

ITK uses an extensible registration framework

Various interchangeable classes exist Relatively easy to “twiddle” the part

you’re interested in while recycling prior work

6

ITK terminologyITK terminology

Fixed image f(x) - stationary in space Moving image m(x) - the fixed image

with an unknown transform applied Goal: recover the transform T(x) which

maps points in f(x) to m(x)

Fixed image f(x) - stationary in space Moving image m(x) - the fixed image

with an unknown transform applied Goal: recover the transform T(x) which

maps points in f(x) to m(x)

7

Registration framework piecesRegistration framework pieces 2 input images, fixed and moving Metric - determines the “fitness” of the

current registration iteration Optimizer - adjusts the transform in an

attempt to improve the metric Interpolator - applies transform to image

and computes sub-pixel values

2 input images, fixed and moving Metric - determines the “fitness” of the

current registration iteration Optimizer - adjusts the transform in an

attempt to improve the metric Interpolator - applies transform to image

and computes sub-pixel values

8

ITK registration flowchartITK registration flowchart

9

ITK’s “Hello world” exampleITK’s “Hello world” example 2D floating point image inputs Please see the software guide (section

8.2) for code specifics I’m going to cover what each piece

does, not look at code per se

2D floating point image inputs Please see the software guide (section

8.2) for code specifics I’m going to cover what each piece

does, not look at code per se

10

Input imagesInput images

2D floating point Floating point avoids loss of precision

problems with integer pixel types

2D floating point Floating point avoids loss of precision

problems with integer pixel types

11

TransformTransform

TranslationTransform Permits translation only in 2D Documentation notes similarity to

AffineTransform, but improved speed (why…?)

TranslationTransform Permits translation only in 2D Documentation notes similarity to

AffineTransform, but improved speed (why…?)

12

MetricMetric

MeanSquaresImageToImageMetric Sum of squared differences between 2

images on a pixel-by-pixel basis A bit naïve Works for 2 images that were acquired

with the same imaging modality

MeanSquaresImageToImageMetric Sum of squared differences between 2

images on a pixel-by-pixel basis A bit naïve Works for 2 images that were acquired

with the same imaging modality

13

OptimizerOptimizer

RegularStepGradientDescent Follows the derivative of the metric Step size depends on rapid changes in

the gradient Step size eventually reaches a user-

defined value that determines convergence

RegularStepGradientDescent Follows the derivative of the metric Step size depends on rapid changes in

the gradient Step size eventually reaches a user-

defined value that determines convergence

14

InterpolatorInterpolator

LinearInterpolateImageFunction Fast and conceptually simple

LinearInterpolateImageFunction Fast and conceptually simple

15

WrapperWrapper

ImageRegistrationMethod Combines all of the previous classes

into a master class

registration->SetMetric( metric );registration->SetOptimizer( optimizer );registration->SetTransform( transform );registration->SetInterpolator( interpolator);

ImageRegistrationMethod Combines all of the previous classes

into a master class

registration->SetMetric( metric );registration->SetOptimizer( optimizer );registration->SetTransform( transform );registration->SetInterpolator( interpolator);

16

Other stepsOther steps

Set the region of the fixed image the registration will operate on (e.g. to ignore bad data)

Initialize the transform Twiddle the optimizer for best

performance*

Set the region of the fixed image the registration will operate on (e.g. to ignore bad data)

Initialize the transform Twiddle the optimizer for best

performance*

*may involve magic

17

Hello world inputHello world input

18

X & Y translation vs. timeX & Y translation vs. time

19

Metric vs. timeMetric vs. time

20

Registration resultsRegistration results

After registration converges/terminates you call GetLastTransformParamters to recover the final transform

For the Hello World example there are 2 parameters, X & Y translation

After registration converges/terminates you call GetLastTransformParamters to recover the final transform

For the Hello World example there are 2 parameters, X & Y translation

21

Double checking resultsDouble checking results

Use ResampleImageFilter to apply the transform for the fixed image

Take the output, compute a difference image with the moving image, examine the results

Good registration results in nothing much to see

Use ResampleImageFilter to apply the transform for the fixed image

Take the output, compute a difference image with the moving image, examine the results

Good registration results in nothing much to see

22

Image comparisonImage comparison

23

Keeping tabs on registrationKeeping tabs on registration Registration is often time consuming It’s nice to know that your algorithm isn’t

just spinning it’s wheels Use the observer mechanism in ITK to

monitor progress You’ve seen this before with filters

Registration is often time consuming It’s nice to know that your algorithm isn’t

just spinning it’s wheels Use the observer mechanism in ITK to

monitor progress You’ve seen this before with filters

24

Observer stepsObserver steps

Write an observer class that will process “iteration” events

Add the observer to the optimizer As a generic note, observers can observe

any class derived from itk::Object

Start registration as usual

Write an observer class that will process “iteration” events

Add the observer to the optimizer As a generic note, observers can observe

any class derived from itk::Object

Start registration as usual

25

Things observers can doThings observers can do

Print debugging info Update GUI Other small management functions Should not do anything too processor

intensive

Print debugging info Update GUI Other small management functions Should not do anything too processor

intensive

26

Multi-modality registrationMulti-modality registration

Remember how I said sum-of-squares difference is relatively naïve?

Mutual information helps overcome this problem

Section 8.4 shows how to implement a simple MI registration

Remember how I said sum-of-squares difference is relatively naïve?

Mutual information helps overcome this problem

Section 8.4 shows how to implement a simple MI registration

27

Notes about the MI exampleNotes about the MI example Significantly, largely the same piece of

code as Hello World Mutual Information is a metric, so we

can keep the optimizer, the interpolator, and so on

Majority of differences are in tweaking the metric, not in rewriting code

Significantly, largely the same piece of code as Hello World

Mutual Information is a metric, so we can keep the optimizer, the interpolator, and so on

Majority of differences are in tweaking the metric, not in rewriting code

28

MI InputsMI Inputs

T1 MRI Proton density MRI

29

MI OutputMI Output

Before After

30

Centered transformsCentered transforms

More natural (arguably) reference frame than having the origin at the corner of the image

Details are not appreciably different from other rigid registrations, see 8.5

More natural (arguably) reference frame than having the origin at the corner of the image

Details are not appreciably different from other rigid registrations, see 8.5

31

An aside: “Twiddling”An aside: “Twiddling”

A common criticism of many/most registration techniques is their number of parameters

A successful registration often depends on a very specific fine-tuning of the algorithm

“Generalized” registration is an open problem

A common criticism of many/most registration techniques is their number of parameters

A successful registration often depends on a very specific fine-tuning of the algorithm

“Generalized” registration is an open problem

32

Multi-Resolution registrationMulti-Resolution registration Useful to think of this as algorithmic

“squinting” Start with something simple and low-res Use low-res registration to seed the

next higher step Eventually run registration at high-res Also called “coarse to fine”

Useful to think of this as algorithmic “squinting”

Start with something simple and low-res Use low-res registration to seed the

next higher step Eventually run registration at high-res Also called “coarse to fine”

33

Multi-resolution schematicMulti-resolution schematic

34

Image pyramidsImage pyramids

35

OptimizationOptimization

Parameter dependency rears its ugly head

You often/usually need to adjust optimizer parameters as you move through the pyramid

You can do this using the Observer mechanism

Parameter dependency rears its ugly head

You often/usually need to adjust optimizer parameters as you move through the pyramid

You can do this using the Observer mechanism

36

Multi-resolution exampleMulti-resolution example

Again, mostly the same code as Hello World

Use MultiResolutionPyramidImage filter to build fixed and moving pyramids

MultiResolutionImageRegistrionMethod is now the overall framework

Again, mostly the same code as Hello World

Use MultiResolutionPyramidImage filter to build fixed and moving pyramids

MultiResolutionImageRegistrionMethod is now the overall framework

37

Benefits of multi-resolutionBenefits of multi-resolution Often faster More tolerant of noise (from “squinting”) Minimizes initialization problems to a

certain extent, though not perfect

Often faster More tolerant of noise (from “squinting”) Minimizes initialization problems to a

certain extent, though not perfect

38

See the software guide for…See the software guide for… Complete list of:

Transforms Optimizers Interpolation methods

You’re encouraged to mix and match!

Complete list of: Transforms Optimizers Interpolation methods

You’re encouraged to mix and match!

39

Deformable registrationDeformable registration

ITK has 2 techniques: Finite element: treat small image regions

as having physical properties that control deformation

Demons: images are assumed to have iso-intensity contours; image deformations occur by pushing on these contours

ITK has 2 techniques: Finite element: treat small image regions

as having physical properties that control deformation

Demons: images are assumed to have iso-intensity contours; image deformations occur by pushing on these contours

40

Model based registrationModel based registration

Build a simplified geometric model from a training set

Identify parameters that control the characteristics of the model

Register the model to a target image to adapt to a particular patient

Build a simplified geometric model from a training set

Identify parameters that control the characteristics of the model

Register the model to a target image to adapt to a particular patient

41

Model based, cont.Model based, cont.

Uses the Spatial Objects framework for representing geometry

Useful because it derives analytical data from the registration process, not just a pixel-to-pixel mapping

Uses the Spatial Objects framework for representing geometry

Useful because it derives analytical data from the registration process, not just a pixel-to-pixel mapping

42

Model-based reg. schematicModel-based reg. schematic

43

Model-based exampleModel-based example

44

Speed issuesSpeed issues

Execution time can vary wildly Optimizer (more naïve = faster) Image dimensionality (fewer = faster) Transform (fewer DOF = faster) Interpolator (less precise = faster)

Execution time can vary wildly Optimizer (more naïve = faster) Image dimensionality (fewer = faster) Transform (fewer DOF = faster) Interpolator (less precise = faster)

45

Take home messagesTake home messages

Exactly what parameters do what is not always obvious, even if you’re familiar with the code

Successful registrations can be something of an art form

Multi-resolution techniques can help Work within the framework!

Exactly what parameters do what is not always obvious, even if you’re familiar with the code

Successful registrations can be something of an art form

Multi-resolution techniques can help Work within the framework!